Impact of the parameter distribution on the predictive quality of metamodels for clinch joint properties

Abstract. The growing significance of lightweight design, reveals drawbacks of conventional joining processes such as welding, which are known to consume a considerable amount of energy. This fosters the use of mechanical joining processes including clinching. However, the lack of universally applicable design methods results in a cost- and time-intensive design process. The utilization of machine learning methods can overcome these drawbacks. To ensure a reliable clinch joint design, inherent uncertainties of the design parameter such as tool deviations need to be considered in the design process. Varying distributions of design parameters, due to changes in the manufacturing process, can lead to high-computational effort in recalculating the resulting clinch joint properties with numerical simulations. Current metamodel-based methods for consideration of inherent uncertainties within the design parameters do not investigate the transferability of metamodels to different distributions of design parameters, which can lead to incorrect predictions. Therefore, this contribution investigates the performance of several metamodels on differently distributed design parameters. The obtained results indicate that metamodels demonstrate the best performance when training and evaluation distributions are identical and that polynomial regression models perform best on disparate distributions, when trained on uniform distributions.